Cheeseman: a travesty of truth

Though I am in sympathy with Cheeseman’s cause of promoting Bayesian inference, I am alienated by his philosophical position. Reading his paper was a little like reading a pamphlet on the need for more science content in school curricula, and finding that it was issued by the Creation Science Council. Let me first indicate the extent of my agreement with Cheeseman’s cause. Like Cheeseman, I felt that McDermott’s “A critique of pure reason” (McDermott 1987) underrated the potential of Bayesian methods for nondeductive inference (as noted in my commentary on that paper (Schubert 1987)). Furthermore, I agree that Bayesian methods are often wrongly maligned for being “too numerical,” or requiring too much information. Like Cheeseman, I believe that Bayesian methods are appropriate for modelling the “fuzziness” (or better, “looseness”) of natural language. Indeed, I have advocated the very same approach to the interpretation of degree adjectives that Cheeseman illustrates with the examples of tall (Schubert 1978). (I also cited some empirical data favouring that approach over others in the literature.) Finally, I agree with Cheeseman that belief modification in knowledge-based systems ought to be founded on an inductive probability theory rather than on current nonmonotonic logics. However, I think that Cheeseman’s advocacy of Bayesian methods is marred by two misapprehensions. The first (and wont) is that he pits Bayesian methods against logical methods as if they were antagonistic alternatives. This imagined antagonism stems largely from misconceptions about truth and belief. Near the end of the paper, he invites logic back on board the AI ship, but only after duly keelhauling it. The second is that he sometimes misrepresents or misapplies Bayesian methods. I will elaborate on these criticisms in turn. I will conclude with remarks on the relation between logic and probability.